이미지: X — 모델·오픈소스 영상 갈무리
Summary
- On August 18, Alibaba's Qwen introduced a local-run demo of its 27B model and a user comparison case via its official X account
- User AJ said that in a Three.js water simulation created with the same prompt, Qwen3.8-27B outperformed Google's Gemini 3.7 Flash
- The test was run with a Q5 quantized version on a single RTX 3090; Qwen3.8-27B was released under Apache 2.0 on the 14th and has since topped Hugging Face's trending chart
- 공개일/라이선스
- 2026-08-14, 아파치 2.0 오픈웨이트 (알리바바 큐원)
- 허깅페이스 트렌딩
- 8월 16일 기준 1위, 원본 다운로드 9만1900건·좋아요 9890건(unsloth GGUF 판 86만8000건)
- 이번 비교 테스트
- 사용자 AJ가 동일 Three.js 물 시뮬레이션 프롬프트로 Qwen3.8-27B와 구글 제미나이 3.7 플래시를 원샷 비교
- 실행 환경
- RTX 3090 GPU, Q5 GGUF 양자화 버전(F8 아님)
- 제3자 코멘트
- 레온 츠베트코프스키는 제미나이 3.7 플래시 가격대를 근거로 MoE 60~70B급으로 추정하며 비교가 공정하다고 평가
- 큐원 공식 발표
- 8월 18일 공식 X 계정이 로컬 실행 데모와 AJ의 비교 게시물을 소개
A 27B model run on a single laptop, reportedly beating Gemini
Alibaba's official Qwen X account posted a short video on August 18. The caption was simple: strong enough to keep pace with frontier models, yet light enough to run on a laptop. A second post that followed minutes later cited a comparison demo shared by user AJ (@ItsmeAjayKV). Under the same prompt and a single-shot condition, Qwen3.8-27B reportedly produced results that surpassed Google's Gemini 3.7 Flash.
Qwen3.8-27B takes the stage again
This model isn't new. Alibaba's Qwen released the Qwen3.8-27B weights on Hugging Face under an Apache 2.0 license back on August 14. It's a dense model with just under 30 billion parameters, and its 4-bit GGUF conversion is only 17.9GB — small enough to fit on a single 24GB-class GPU. Two days later, on August 16, the model reportedly topped Hugging Face's trending chart. The original weights recorded 91,900 downloads and 9,890 likes, while a GGUF conversion made by community developer unsloth pulled in far more downloads — 868,000. This latest post is essentially a follow-up promotional push just days after that milestone.
A user-submitted water simulation comparison
The test AJ shared involved simulating water waves using Three.js — a JavaScript library for rendering 3D graphics in web browsers. He ran the same prompt once each on Qwen3.8-27B and Google's Gemini 3.7 Flash and posted the results side by side. In AJ's comparison post, he wrote, "the difference is hugeee." He specifically noted that he ran the test on a single RTX 3090 using a Q5 quantized version rather than F8. Quantization — a compression method that reduces a model's computational precision to cut down on size and compute — can cause greater performance loss at lower tiers, making it notable that the result still came out ahead under those conditions.
Leon Cvetkovski commented on the post, estimating — based on Gemini 3.7 Flash's pricing structure — that it is likely a mixture-of-experts (MoE) model in the 60–70B range, and suggesting the comparison isn't drastically unfair if that's the case. However, this is only a third-party estimate; Google has not confirmed disclosing the architecture of Gemini 3.7 Flash.
The exact prompt used
Here is the full prompt AJ used identically for both models:
"Create a Three.js water simulation. Requirements: Plane mesh with 512x512 subdivisions. Wave propagation simulation. Mouse click creates ripples. Multiple interacting waves. Real-time lighting. Everything in a single HTML file. No external libraries besides Three.js."
In short, the task requires building a water simulation in Three.js with a 512x512-subdivided plane mesh, wave propagation, ripples triggered by mouse clicks, multiple interacting waves, and real-time lighting — all within a single HTML file using no libraries other than Three.js. Pasting this prompt directly into a local chat UI or API allows anyone to reproduce a comparison under the same conditions.
Comparing the two models by spec
| Item | Qwen3.8-27B | Gemini 3.7 Flash |
|---|---|---|
| Developer | Alibaba Qwen | Google DeepMind |
| Disclosed architecture | 27B dense, Apache 2.0 open-weight | Undisclosed (estimated 60–70B MoE based on pricing) |
| Environment used in this test | RTX 3090, Q5 GGUF | Cloud API |
| Recent history | Released 8/14, topped Hugging Face trending 8/16 | No confirmed background |
Editor's take
This post itself isn't a new announcement — it's a promotional push. But the timing is notable. Qwen3.8-27B has now gone through four rounds of attention in under two weeks: release (8/14) → topping the trending chart (8/16) → criticism over excessive reasoning (same week) → this latest citation of a user comparison (8/18). Alibaba's continued re-citing of community demos through its official account reads as a strategy to prove the model's power not through benchmark tables, but through firsthand "here's what happened on my machine" testimonials. Benchmark scores are not immune to manipulation controversies, but a demo reproducible on a single RTX 3090 carries a different kind of persuasive weight — anyone can verify it.
For anyone who has worked with local LLMs in practice, this pattern feels familiar. A 27B-class model running on a single 24GB GPU means access to code-generation quality once expected only from 70B+ models, at a fraction of the hardware cost. That said, results from lower quantization tiers like Q5 aren't always stable. Without running the same prompt ten times to check variance, it's premature to greenlight production adoption based on a single "one-shot win" demo.
The takeaway for domestic dev teams is clear: for tasks where a single code output produces immediately visible results — frontend prototypes or 3D visualizations, for example — a single RTX 3090- or 4090-class GPU paired with the Qwen3.8-27B GGUF version lets you run similar experiments right now, with no cloud API costs. But production code or security-sensitive work still needs broader validation. In the coming weeks, expect more of these community-driven one-shot comparisons to extend to other coding and visualization tasks, with Alibaba likely to keep re-citing such cases through its official account.
